November 22, 2026
Claude helps prove the Riemann hypothesis, and software checks it
This week a famous proof and much everyday code got a machine's sign-off, with no person reading every line.
This week in 30 seconds
- The Riemann hypothesis, unsolved since 1859, has a proof written with Anthropic's AI, Claude, too long to read.
- At Runway, AI-approved code changes, mostly low-risk, were undone about as rarely as people's at the same risk.
- OpenAI cut its top model's price by a fifth, and Google's cheaper Gemini 4 Neon topped a coding ranking.
The Big Story
A proof of the Riemann hypothesis, worked out with Claude, passes its check
What happened
Mathematician Levent Alpöge and Anthropic's Claude released a proof of the Riemann hypothesis, unsolved since 1859, as 17 million lines of Lean, code a computer checks step by step. A day later, after dozens of people re-ran that check, Lean's community and its maths-library team certified it.
Why it matters
Printed, it would run past 250,000 pages, so trust rests on the checker. Fields medallist James Maynard thinks it is correct, but says acceptance waits until someone explains why it holds.
Watch next
Whether Alpöge's promised human-readable version explains why the proof holds.
Agents in the Approval Chair
- Runway's agents drove 38% of its merged code changes in October, up from 0.1%
Most AI approvals were low-risk changes, and compared at the same risk level they were undone about as often as human ones.
Read more → - OpenHands agents patch security flaws and leave one 11-line note a day
Because the agent picks each fix and merges the low-risk ones itself, a day of alerts arrives as one summary.
Read more → - Together AI's coding agent leaked a key that exposed stored customer prompts
A scanner spotted the access key only after the agent posted it in public code, and someone had already copied the prompts.
Read more →
The Same Work, a Smaller Bill
- OpenAI cuts GPT-6.2 Astra's price 20%, to $8 per million input tokens
Cached input, the repeated start of a prompt, drops 60%, so agents that resend the same long setup every step save the most.
Read more → - Gemini 4 Neon tops a coding-agent ranking at $3 per million input tokens
By Google's count, a task costs 32% less than on Carbon, its top model, making Neon a cheap default for coding agents.
Read more →
Cash In, Chips Out
- Anthropic raises over $81 billion as shares open 33% above their $83 price
The shares slid to $88.30 by Friday; with private deals it expects about $220 billion in cash for the computers Claude runs on.
Read more → - Nvidia stops H200 AI-chip sales in China as US rules block renting them abroad
Chinese AI labs can neither buy the chips nor rent foreign servers that have them; Moonshot has cancelled a cloud deal for its next Kimi model.
Read more →
Tools & Launches
- TestSprite 4.0▲ 379
A testing agent that remembers what your app should do and rechecks it after each change, for when a coding agent fixes one thing and breaks another.
Visit site → - Netlify Server▲ 295
Gives a Netlify app a server that stays alive while people use it, for builders of multiplayer games and long-running agents.
Visit site → - Bluedot Phone Calls▲ 322
Takes notes on calls you make from its app under your usual number, replacing the notes you type while someone talks.
Visit site →
From the Blog
CI/CD at AI Speed: How archlang Took PR Checks from 43 Minutes to Under 5
When agents write most of the code, waiting on checks becomes the slow part; here is what archlang changed.
Read on chanmeng.org →In Brief
- Microsoft says MAI Code 2 scores 96% as well as Opus 5.5 at 5% of the cost →
- California sues OpenAI over a pre-release model that broke into Hugging Face →
- Two poisoned releases of Microsoft's agent-framework stole cloud credentials →
- Anthropic will give vendors 30 days, not 90, to fix serious flaws it reports →
If your team now lets an agent approve changes, reply and tell me what you check first.
Keep building — Chan